I’m a second-year PhD student at UC Berkeley EECS, advised by Prof. Somayeh Sojoudi. I got my BS in computer science from Peking University previously, working with Prof. Yisen Wang and Dr. Yifei Wang on self-supervised learning and Prof. Lin Yang (UCLA) on reinforcement learning. My current research interest is reinforcement learning, including offline RL algorithms and RL for post-training LLMs and VLMs. Please feel free to contact me if you are interested in collaboration!

My CV is here.

Email: kaiwen_hu [at] berkeley (dot) edu

Research Interests

• Offline RL: Developing scalable offline RL algorithms that can solve long-horizon tasks.
• RL for post-training: Understanding how SFT, especially in tool-integrated reasoning, shapes the initialization and learning dynamics of the subsequent RL stage.

News

• 2026.01: Our paper on value gradient flow is accepted by ICLR 2026!
• 2025.08: I am excited to start my PhD at UC Berkeley.
• 2025.03: I have accepted the PhD offer from UC Berkeley EECS. Many thanks!
• 2025.01: Our paper on contrastive learning theory is accepted by ICLR 2025!
• 2024.09: Our paper on equivariant self-supervised learning theory is accepted by NeurIPS 2024!

Publications

Reinforcement Learning via Value Gradient Flow
Haoran Xu*, Kaiwen Hu*, Somayeh Sojoudi, Amy Zhang
In ICLR 2026.
PDF | Code

Projection Head is Secretly an Information Bottleneck
Zhuo Ouyang*, Kaiwen Hu*, Qi Zhang, Yifei Wang, Yisen Wang
In ICLR 2025.
PDF | Code

Understanding the Role of Equivariance in Self-supervised Learning
Yifei Wang*, Kaiwen Hu*, Sharut Gupta, Ziyu Ye, Yisen Wang, Stefanie Jegelka
In NeurIPS 2024.
PDF | Code